NASGym

A proof-of-concept OpenAI Gym environment for Neural Architecture Search (NAS).

Under development: The code and documentation of this repository may contain minor bugs.

Overview

This is a python package developed for the research project Learning to reinforcement learn for Neural Architecture Search.

The environment is fully compatible with the OpenAI baselines. It implements the RL steps for NAS, using the Neural Structure Code (NSC) of BlockQNN: Efficient Block-wise Neural Network Architecture Generation to encode the networks and make architectural changes. Under this setting, a Neural Network (i.e. the state for the reinforcement learning agent) is modeled as a list of NSCs, an action is the addition of a layer to the network, and the reward is the accuracy after the early-stop training. The datasets considered so far are the CIFAR-10 dataset (default) and the meta-dataset (it has to be downloaded as specified in the original repository).

Highlights

The code structure allows you to:

• Set the main training variables in a config.ini file (see the resources directory).
• Easily change the layers to consider and their hyperparameters in a yml file (see the resources directory).
• Create customized TensorFlow estimators to plug-in your performance estimation strategy.
• Store a local database of the experiments (CSV file by default, with an option to create your own DB interfaces).
• Create your own dataset handlers to input the TensorFlow estimators.

Installation

The recommended way to install the package is in editable mode:

cd ${GIT_STORAGE}/nas-env
git checkout develop
pip install -e .

Example

To see an example of how to use this environment after installation, check the scripts in the nas-dmlr repo.

Future plans

• Architectural changes to the source code to ease plugging-in the different NAS elements, i.e. the performance estimation strategy and the search space.
• Generate documentation
• Publish the package in PiPy.

Contributing

So far I am the only contributor to this project, but I would like to improve the design to allow for even easier integration of the different NAS elements. Please feel free to join me in this quest!